fix(embedding_linker) Mention_Mask fix and CuiEmbed by longest name

[adam-sutton-1992]

Two fixes here leading to a performance upgrade.

1. Provide mention masks for embed_cui and name methods. I previously (and erroneously) thought that mean pooling would be the correct attention for names and cui embeddings, as that's the only thing there. However it does contain the and special tokens as part of it. Throwing off scoring - if you detected say "glucose", embed with no context and just scored that glucose vs an emebdded "glucose" the scoring could be quite far off. You would expect a similarity of 1, but this would return 0.875 here. Using mention masks on the existing mention mask method was in two for loops. Doing this for 3 million names and 800k cuis was quite long, so it's now a tensor based operation.
2. Embed CUIs via their longest name rather than their preferred name. Which solves issues with disambiguation. Another example to explain this; the name "glucose" is ambiguous. It could be a measurement (snomed: 33747003), or a substance (67079006). When embedding via preferred names; the preferred name for the measurement version is "Glucose measurement, blood", the subtance version is... "glucose". The model will always choose substance over measurement due to cosine similarities. The longest names however makes it an easier task - improving performance.

These changes performance in metrics as such to:

Epoch: 0, Prec: 0.2862910530810839, Rec: 0.7398836198247609, F1: 0.4128382160850905.

An improvement of previous similar recalls which were around 0.66.